Deep toxigraph an integrated graph neural network frome work for toxicity prediction

Abstract

The problem of predicting the toxicity of chemical compounds is among the most important problems in the field of drug discovery. Most conventional approaches to modeling the relationship between the chemical structures of compounds and their toxicities encode the structures of those chemicals as vectors of features, thereby discarding information about the relationship between the chemicals and their structures during the encoding process. As a result, the models trained using such encodings tend to have limited predictive power when applied to diverse libraries of chemical compounds. In this work, DeepToxiGraph is introduced as a deep learning framework for the task of toxicity prediction that uses several different graph neural network architectures to encode chemical compounds and predict their toxicity. Through a series of ablation studies, each of which eliminates a certain feature of the models, the relative importance of various features of graph neural networks is studied. Based on these studies, a novel hybrid model that combines graph, fingerprint, and descriptor representations is proposed. The hybrid model achieves an R² of 0.6307 and an RMSE of 0.4161 on the mouse intraperitoneal LD₅₀ test set. This entire framework is compiled into a web site that allows chemistry scientists to predict the toxicity of chemical compounds by simply inputting their SMILES notations, without the need to learn any programming language or skills